如何将未见过的测试数据传入交叉验证函数并预测标签？

Got it, let's walk through how to adapt your code to predict labels on that unseen test data. First, a quick clarification: your existing cross_val() function is for evaluating model performance on splits of your training data, not generating predictions on external test sets. We’ll keep that for validation, but add a separate workflow to train a final model and make predictions.

Step 1: Keep Cross-Validation for Performance Checks #

Your cross-validation code is solid for assessing how well your model generalizes. We’ll leave that as-is to confirm performance before moving to predictions.

Step 2: Add a Training & Prediction Workflow #

Here’s the updated code that combines your existing setup with the logic to predict on X2:

import pandas as pd
from sklearn.model_selection import KFold, cross_val_score
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.pipeline import make_pipeline
from sklearn import preprocessing
from sklearn.svm import SVC

# Load your datasets
df = pd.read_csv('./output/csv_sanitized_16_.csv', dtype=str)
X = df['description_plus']
y = df['category_id']

df_2 = pd.read_csv('./output/csv_sanitized_2.csv', dtype=str)
X2 = df_2['description_plus']

# Your original cross-validation function (unchanged)
def cross_val():
    cv = KFold(n_splits=20)
    vectorizer = TfidfVectorizer(sublinear_tf=True, max_df=0.5, stop_words='english')
    X_train = vectorizer.fit_transform(X)
    clf = make_pipeline(preprocessing.StandardScaler(with_mean=False), SVC(C=1))
    scores = cross_val_score(clf, X_train, y, cv=cv)
    print(scores)
    print("Accuracy: %0.2f (+/- %0.2f)" % (scores.mean(), scores.std() * 2))

# Run cross-validation to validate model performance
cross_val()

# Function to train on full training data and predict on unseen test data
def train_and_predict_unseen():
    # Fit the TF-IDF vectorizer ONLY on training data (critical!)
    vectorizer = TfidfVectorizer(sublinear_tf=True, max_df=0.5, stop_words='english')
    X_train_transformed = vectorizer.fit_transform(X)
    
    # Build and train the full pipeline on all training data
    clf = make_pipeline(preprocessing.StandardScaler(with_mean=False), SVC(C=1))
    clf.fit(X_train_transformed, y)
    
    # Transform test data using the pre-fitted vectorizer (don't re-fit!)
    X2_transformed = vectorizer.transform(X2)
    
    # Generate predictions for the unseen data
    predicted_labels = clf.predict(X2_transformed)
    
    # Optional: Attach predictions back to your test dataframe and save
    df_2['predicted_category_id'] = predicted_labels
    df_2.to_csv('./output/test_with_predictions.csv', index=False)
    
    return predicted_labels

# Run the prediction function
test_predictions = train_and_predict_unseen()
print("Generated predictions for unseen test data:", test_predictions)

Key Things to Note #

• Don’t fit on test data: Always use vectorizer.transform(X2) for test data, not fit_transform. This ensures we use the same vocabulary and TF-IDF weights learned from the training set—using fit_transform on test data would leak information and invalidate your predictions.
• Cross-validation ≠ prediction: Cross-validation is for estimating performance. To get real predictions on new data, you need to train a model on the entire training dataset (since cross-val uses splits of training data for testing, not external data).
• Pipeline consistency: Using the same pipeline (vectorizer → scaler → SVC) for both cross-validation and final training ensures your preprocessing steps are applied the same way everywhere.

内容的提问来源于stack exchange，提问作者neelmeg